Hybrid deep boosting ensembles for histopathological breast cancer classification

Purpose Breast cancer (BC) is the most common diagnosed cancer type and one of the top leading causes of death in women worldwide. This paper aims to investigate ensemble learning and transfer learning for binary classification of BC histological images over the four-magnification factor (MF) values of the BreakHis dataset: 40X, 100X, 200X, and 400X. Methods The proposed homogeneous ensembles are implemented using a hybrid architecture that combines: (1) three of the most recent deep learning (DL) techniques for feature extraction: DenseNet_201, MobileNet_V2, and Inception_V3, and (2) four of the most popular boosting methods for classification: AdaBoost (ADB), Gradient Boosting Machine (GBM), LightGBM (LGBM) and XGBoost (XGB) with Decision Tree (DT) as a base learner. The study evaluated and compared: (1) a set of boosting ensembles designed with the same hybrid architecture and different number of trees (50, 100, 150 and 200); (2) different boosting methods, and (3) the single DT classifier with the best boosting ensembles. The empirical evaluations used: four classification performance criteria (accuracy, recall, precision and F1-score), the fivefold cross-validation, Scott Knott statistical test to select the best cluster of the outperforming models, and Borda Count voting system to rank the best performing ones. Results The best boosting ensemble achieved an accuracy value of 92.52% and it was constructed using XGB with 200 trees and Inception_V3 as feature extractor (FE). Conclusions The results showed the potential of combining DL techniques for feature extraction and boosting ensembles to classify BC in malignant and benign tumors.